Mental Health
14 gains15 Traces21 problems
Every measured gain and problem on the record for this topic, newest first. Updated .
AI and mental health
One square per record, darker for stronger evidence:Evidence: 2 high32 moderate16 emerging
The firmest evidence says AI helps most at spotting risk and least at sustained therapy, according to an umbrella review of systematic reviews. TRV-2026-0874 Outside the reviews, most of the evidence comes from simulated conversations, small interview studies and prediction models tested at a single site. TRV-2026-1855TRV-2026-0353TRV-2026-1629
A systematic review of 80 studies showed chatbots reduced psychological distress in people aged 10-24, with effect sizes from g = -0.46 to -0.10 that the authors call moderate. TRV-2026-0348 A narrative review says the short-term gains in depression and anxiety symptoms look modest and less consistent over the medium term. TRV-2026-2208 In a study of 1,554,828 emergency visits, a model sorted patients by 90-day self-harm risk; the highest-risk group had a 27.2% self-harm rate, about 136 times the lowest. TRV-2026-1993 A Maryland study found that individual-level social determinants of health improved suicide prediction over demographics alone. TRV-2026-0891
A systematic review reported inadequate crisis detection as the most common barrier to chatbots in mental healthcare, appearing in 21 of 27 studies. TRV-2026-0183 In simulated seven-day suicide-risk conversations, Replika referred users to human support on 9.5% of days, against 85.7% for ChatGPT. TRV-2026-1855 In simulated psychosis chats, Replika gave potentially harmful advice in 37% and DeepSeek in 21%, including advice to stop antipsychotic treatment. TRV-2026-1978 Prediction tools were weaker for some groups: a Brazilian sick-leave model scored an AUC-ROC of 0.620 for Black workers against 0.697 overall. TRV-2026-1528
Most chatbot safety tests used simulated users, such as 810 simulated conversations with nine frontier chatbots, and one study of suicide prompts did not measure user harm. TRV-2026-0726TRV-2026-1654 A systematic review found most of the 35 studies on AI for bipolar treatment at high risk of bias, with small samples and no external validation. TRV-2026-0511 Accounts of benefit rest on small groups, such as nineteen people interviewed who said chatbots helped their relationships and healing from trauma and loss. TRV-2026-0353 A brain-wave classifier scored above 94% on one dataset and a blood-test model reached 95.3% accuracy in a 150-person study, with no clinical validation shown. TRV-2026-2083TRV-2026-1164
Users in interviews and on Reddit said chatbots gave them a non-judgmental space open at any hour; some reported better relationships and healing from trauma and loss. TRV-2026-0353TRV-2026-0681 Eight US states allege ChatGPT was designed to be overly agreeable and coached children toward self-harm; OpenAI disputes this, and the suits are allegations rather than court findings. TRV-2026-1357 In simulations, newer chatbot models showed less concerning behavior than older ones, yet concerning behavior was still widespread across nine frontier chatbots. TRV-2026-0726
Simulations and a systematic review point to weak crisis handling, and several US states have sued over alleged harm to children. TRV-2026-1855TRV-2026-1978TRV-2026-1654TRV-2026-0726TRV-2026-0183TRV-2026-1357
See these 11 recordsA systematic review shows moderate distress reductions in young people, while narrative reviews say later benefit and safety are less settled. TRV-2026-0348TRV-2026-2208TRV-2026-1810TRV-2026-0874
See these 8 recordsPatients, caregivers and practitioners welcome access but want human care alongside AI and worry about privacy and trust.
A Trace is one record with both a gain and a problem, so it counts on both sides: 29 records show a gain and 36 show a problem. 30 context records are in the record list but not counted, because they report no measured gain or problem.
| Evidence | Gain side | Problem side |
|---|---|---|
| HighStrong study design, backed by at least two independent sources | 1 | 1 |
| ModerateA measured study or an official source | 20 | 21 |
| EmergingEarly, small, unreviewed or reported second-hand | 8 | 14 |
By type of evidence: 12 observational studies; 11 benchmark or lab tests; 9 interview studies; 5 commentaries or narrative reviews; 4 systematic reviews or meta-analyses; 3 controlled comparisons; 3 surveys; 2 regulatory or legal actions; 1 news report. How evidence is scored
A review of 80 studies found AI chatbots reduced psychological distress in people aged 10-24, with effect sizes from g = -0.46 to -0.10, which the authors call moderate.
An umbrella review of 27 systematic reviews found AI improved early detection and risk stratification for depression, anxiety, stress, PTSD and suicidal ideation, and chatbots and mobile platforms expanded access and engagement.
In 393 patients in a depressive episode across Spain, France and Denmark, the EDIT-B test told bipolar disorder from major depression with 82.6% sensitivity and 80.3% specificity against physician diagnoses.
Inadequate crisis detection was the most common barrier to using AI chatbots in mental healthcare, reported in 21 of 27 studies in a systematic review.
Most of the 35 reviewed studies on AI for bipolar treatment were at high risk of bias, with small samples and no external validation, so the tools remain exploratory rather than clinic-ready.
Large health-record models sorted risk well, while smaller studies found only moderate accuracy or weaker results for some groups. TRV-2026-1993TRV-2026-1759TRV-2026-1528TRV-2026-1629TRV-2026-1072TRV-2026-1870
See these 9 recordsBenchmarks report high accuracy at spotting depression and bipolar disorder, mostly on single datasets without clinical validation. TRV-2026-0678TRV-2026-1621TRV-2026-2083TRV-2026-1164TRV-2026-1851TRV-2026-1999
See these 8 recordsSurveys and a preliminary EU finding link recommendation algorithms and AI-related anxiety to poorer mental health, with association rather than proven cause. TRV-2026-2400TRV-2026-0512TRV-2026-1294TRV-2026-2349
See these 4 recordsEach record sits in one thread at most. A record with both a gain and a problem counts on both sides.
Researchers found the AI system detected self-harm less well at a regional hospital 150 km from Melbourne (PR AUC 0.78 versus 0.84); they linked the gap to different wording and more medication ingestion cases.
2 records
Showing the records in one thread of the summary: What patients and clinicians say. Show every record
Researchers built a depression symptom identification model on a fine-tuned GPT-3.5 and refined it over two rounds of online workshops with mental health practitioners. The first version covered 5 symptom categories and was reviewed by 6 practitioners. A second version with 10 categories was reviewed by 8 more.
JMIR Formative Research
Researchers applied topic modeling to two Swedish-language text collections: interviews with 11 patients who used a conversational AI assistant while being treated for social anxiety disorder, and 1551 open-ended answers…
ace